Accelerator Error Detection via Utilization-Based Parameter Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional error detection methods for accelerator devices are inefficient due to fixed detection parameters, leading to misreporting of errors during heavy workloads and delayed error detection, which negatively impact system performance and user experience.
Innovation Solution
A workload-aware error detection method that adjusts detection parameters such as timeout value and frequency based on the current utilization rate of the accelerator device, using a mapping table established through experimentation or machine learning, to improve error detection accuracy and minimize system performance impact.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If fixed detection parameters are used for error detection, then the detection method is simple, but error detection accuracy deteriorates during heavy workloads
Solution Approach 1:
The patent implements dynamic adjustment of detection parameters (timeout values, detection frequencies) based on the accelerator device's utilization rate. The system transitions from static fixed parameters to dynamic parameters that adapt to workload conditions, resolving the contradiction between simplicity and accuracy by automating the adjustment process based on real-time device state
Solution Approach 2:
The patent changes detection parameters (timeout values, detection intervals) according to the accelerator device's utilization rate. When utilization is high, timeout values are increased and detection frequency is reduced; when utilization is low, timeout values are decreased and detection frequency is increased. This parameter adaptation resolves the contradiction by optimizing detection accuracy for different operational conditions
2Ease of operation
If fixed detection parameters are used, then the system is easy to operate, but misreporting of errors occurs during heavy workloads
Solution Approach 1:
The system automatically monitors its own utilization rate and self-adjusts detection parameters without external intervention. The accelerator device or management system autonomously determines when to increase timeout values or reduce detection frequency based on current workload, eliminating the need for manual configuration while preventing false error reports during heavy operations
Solution Approach 2:
The system implements a feedback loop where detection parameters are continuously adjusted based on real-time utilization rate monitoring. The utilization rate feedback triggers automatic parameter modification to prevent misreporting, maintaining reliability while keeping the system easy to operate through automated closed-loop control
3Loss of time
If detection frequency is increased to detect errors earlier, then error detection timeliness is improved, but system performance deteriorates
Solution Approach 1:
The patent dynamically adjusts detection frequency based on utilization rate. During low-utilization periods, detection frequency is increased to improve error detection timeliness. During high-utilization periods, detection frequency is reduced to minimize performance impact. This dynamic approach resolves the contradiction by optimizing the trade-off between detection speed and system performance according to actual workload conditions
4Speed
If timeout value is decreased to detect errors faster, then error detection speed is improved, but false error reports increase during heavy workloads
Solution Approach 1:
The patent changes the timeout parameter based on utilization rate. During low-utilization periods, shorter timeout values enable faster error detection. During high-utilization periods, longer timeout values prevent false error reports by accounting for delayed responses due to heavy workload. This parameter adaptation resolves the contradiction between detection speed and reporting accuracy
Data Source
AI summary
Embodiments of the present disclosure relate to a method, device and computer program product for error management. The method comprises obtaining a utilization rate of an accelerator device in accelerator resources, and the accelerator resources comprise one or more accelerator devices. The method further comprises adjusting a detection parameter for detecting an error in the accelerator device based on the utilization rate of the accelerator device, and performing error detection for the accelerator device based on the detection parameter.